Show HN: Fluxnation, the Fastest FLUX.1 Execution Kernel (github.com)

A new tool called Fluxnation, which features a fused CUDA kernel for FLUX.1 image generation optimized for NVIDIA's RTX 4090, has been released, achieving remarkable image generation speeds—approximately 9.3 seconds for 20 steps at a resolution of 1024×1024. This custom kernel dramatically improves performance by integrating multiple technologies into a single ComfyUI node, including an FP8 fused CUDA kernel that replaces several operations with a single call, reducing computational overhead, while also implementing spike attention and step caching to optimize resource usage during image processing. This advancement is significant for the AI/ML community as it addresses the long-standing challenge of efficiently processing high-resolution images, making the workflow faster and more practical for real-time applications. The integration of neuromorphic block-sparse Triton attention allows for significant reductions in computational load—by up to 90% during specific steps—without sacrificing output quality. For developers and researchers utilizing ComfyUI, Fluxnation offers a streamlined process to enhance their projects with this innovative approach to deep learning, thus pushing the boundaries of what's achievable in rapid image generation.